When is a machine learning model said to be overfitted?

A machine learning model is considered overfitted when it performs exceptionally well on the training data but poorly on new, unseen data. Overfitting occurs when a model learns not just the underlying patterns in the training data but also the noise and anomalies that are specific to that dataset. This level of detail can lead the model to make predictions that are inaccurate when faced with data that it hasn't encountered before.

In practical terms, overfitted models have high variance; while they may show impressive accuracy on training datasets, their performance deteriorates significantly when tested with external data. This indicates that the model has not been able to generalize effectively, which is the ultimate goal of a successful machine learning application.

The other options present scenarios that are contrary to the concept of overfitting. A model that generalizes well to new data implies it is performing effectively and has not overfitted. Capturing only underlying patterns signifies a model that is likely well-balanced and not overfitting. Finally, using alternative data sources does not inherently relate to overfitting unless the model fails to accommodate those sources adequately; however, it does not directly indicate overfitting in itself.